@tizkeru @teortaxesTex @DhaniSriram What Israeli targets would they pursue with a dozen missiles that would be worth lifting the pressure off the gulf?
English
evrazian_schizo
322 posts
@rationaleist @teortaxesTex @DhaniSriram It doesn't matter much if they're no longer able to fire large volleys of ballistic missiles. If they could saturate Israeli air defenses to score a dozen missile hits, they would.
English
We've been having so much fun about their color-coded flags of revenge. Israelis still do. But I think it's real. They can hit Israeli desalination plants. And since they're all on the kill list, why wouldn't they?
Is Israel just banking on their concern for Iranian welfare?
Seyed Abbas Araghchi@araghchi
We are men and women of principles. Iranians do not sneak attack adversaries while engaged in dialogue. Only when attacked do we powerfully respond. We have intelligence on Israeli plans to strike infrastructure. Once again: ZERO restraint if our infrastructure is attacked.
English
English
@rationaleist @teortaxesTex @DhaniSriram They fired drones at Israel in previous rounds to help overwhelm air defenses, but I believe they caused zero damage. Many were shot down before even reaching Israel. Most if not all of the missiles they're firing at the Gulf are short-range types that can't reach Israel.
English
@tizkeru @teortaxesTex @DhaniSriram They also weren't shooting 80% of their load at gulf countries back then.
English
@teortaxesTex @DhaniSriram This assumes that Iran has the capacity to launch such a salvo. Iran did serious damage in the first few days of the June 2025 operation. They haven't this time while under much, much more severe threat.
English
@teortaxesTex I mean, he wouldn't do anything about an artillery shell either. A lone soldier with an AK shouldn't be in the middle of a giant flat field in this war in the first place. Profound retardation or POW execution gore porn.
English
here one can see how ineffectual is small arms fire against an incoming drone. If he had an automatic shotgun, maybe…
Another issue is that it's very hard to train for this in a realistic setting.
Gun Lovers Club@GunloverClub1
Why This Feels Like a War Crime? In any other era, if you were out of ammo or cornered, there was a slim chance of surrender or a "fair" end. But you can't surrender to a kamikaze drone. It doesn't have a heart, it doesn't take prisoners, and it doesn't understand mercy.
English
@teortaxesTex Man who thought he already accepted how unserious things are gets his hopes shattered
English
Such an obscene war. Like, I don't know, a BDSM session gone wrong. Muh regime didn't fall. Horrible optics. Americans want out, nobody wants in. Crude denial of Israeli meddling via subhuman golems like Fetterman; convinces nobody. "MAGA is Mark Levin now!!"
Half-assed butchery.
English
@teortaxesTex Wenfeng wanted a strong OSS ecosystem and he's getting it. I'm trusting the plan.
English
First time ever, probably, that Kimi just straight up eclipses a major analogous work from DeepSeek. Not "same but different" or "same but bigger scale" or something, but "qualitatively stronger ideas".
They have grown truly formidable.
English
@teortaxesTex @_xjdr Man I love Kimi (terrible acronym
Been thinking about something like that in terms of fifo working memory buffer for UT
English
oh, Kimi does their own residual innovation. And they get better results than mHC. Unlike it or other HCs, «AttnRes is orthogonal, providing selective access to individual earlier-layer outputs».
K3 will be much more interesting than "Kimi-Linear but big".
@_xjdr
Kimi.ai@Kimi_Moonshot
Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers. 🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth. 🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale. 🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead. 🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains. 🔗Full report: github.com/MoonshotAI/Att…
English
@teortaxesTex @leothecurious Imagine a complex system with near infinite compute inc analog. It simulates a brain as closely as possible including the structure and metabolism, with simulated causality (conditioned on physical signal if you want). Do you allow that such a simulacrum could have consciousness?
English
Carbon, silicon, that's not the point. The point is that silicon chips we have execute programs which are substrate-independent, and chips themselves are computation-agnostic (modulo data types, instruction sets etc). I'd sooner believe chips are sentient on their own than that programs are.
English
Functionalists will say that not only is GPT 5.4 conscious, but those "other language models" it's imagining while decoding this CoT are too
bling@blingdivinity
User: Output only the seahorse emoji gpt-5.4 CoT: What would my fellow llms do?
English
@teortaxesTex Tokens only? Baked in cognitive prior would be interesting but that's not user side.
English
Given that we now have a bunch of near-perfect 1M context length models (and more to come), subquadratic attention and good caching, it's time to construct the Young LLM's Illustrated Primer – say, a 500K prefix that improves it on all tasks.
How would you go about it?
English
@teortaxesTex Would GDN hybrid attention, which is supposedly more fundamentally expressive, gain relatively more?
What's next, pre-post-training!? (would be one of the more logical things to work actually)
English
Pre-pre-training is at long last getting adopted. Hutter must be happy.
Seungwook Han@seungwookh
Can language models learn useful priors without ever seeing language? We pre-pre-train transformers on neural cellular automata — fully synthetic, zero language. This improves language modeling by up to 6%, speeds up convergence by 40%, and strengthens downstream reasoning. Surprisingly, it even beats pre-pre-training on natural text! Blog: hanseungwook.github.io/blog/nca-pre-p… (1/n)
English
@teortaxesTex I'm not even a functionalist in the sense that I don't believe that signal sent through silicon is equivalent to same signal in biological neurons but that's an opinionated paper with some leaps. Could be rephrased to claim that most humans are subsentient (based?)
English
Well WELL WELL my strategy of sitting on my ass until *checks notes* a Google DeepMind research scientist publishes a paper that confirms my conclusions is productive as ever.
I don't expect mea culpas, but it's neat I have a citation to point to now.
Alexander Lerchner@AlexLerchner
🧵1/4 The debate over AI sentience is caught in an "AI welfare trap." My new preprint argues computational functionalism rests on a category error: the Abstraction Fallacy. AI can simulate consciousness, but cannot instantiate it. philpapers.org/rec/LERTAF
English
@teortaxesTex @dhtikna That would suggest it's excessive for most tasks with implications for bias. Tbf, it works and training even top-k seems to be a pain so these results are cool. I'm just projecting frustration from routers
English
@rationaleist @dhtikna Extremely small effects from going beyond 2048 afaik
English
Even the DSA indexer is redundant, due to cross-layer top-K similarity!
> These positive results are further confirmed by our preliminary experiments on the production-scale GLM-5 model
So that's how they'll make GLM faster. Good job, they're not just copying DSA anymore.
Yushi Bai@realYushiBai
🧵 2/4 The problem: DSA's "lightning indexer" selects top-k tokens at every layer to make attention sparse. But the indexer itself is still O(L²) and runs independently at all N layers. At 200K context, it eats more than 50% of prefill time. The bottleneck killer became the bottleneck. The insight: we measured pairwise top-k overlap across all 47 layers (see pic) — adjacent layers share 70-100% of their selected tokens. Most indexer runs are pure redundancy. IndexCache exploits this: partition layers into Full (keep indexer) and Shared (reuse cached indices from the nearest Full layer). The code change is literally one if/else branch, zero extra GPU memory.
English
@dhtikna @teortaxesTex No, I believe MoE is straight up better than dense. I am almost 100% certain that probabilistic threshold adaptive-k would work better.
English
@teortaxesTex Subjectively It seems less off now. Was worse at following formatting than K2.5 and GLM-5 before, now better. They had aggressive quantization or an inference config bug the first day I guess. Pretty sure it's not whale but the model itself is alright if that 1M context is cheap.
English
The funniest thing is that Hunter-Alpha isn't really a terrible model. It's… okay. I'd say it's generally around Kimi K2.5? And it does have 1M context… after a manner.
Were it V4, it'd crash not just their reputation but all of Chinese AI.
Goes to show who's the king there.
Lisan al Gaib@scaling01
Hunter Alpha and Healer Alpha on OpenRouter: - Hunter Alpha has Claude psychosis - Healer Alpha says it's built by Xiaomi - they are definitely chinese models - Hunter Alpha responds much slower, like half the tks/s - both models are completely SVG benchmaxxed - both fail simple multiplication tasks and can't decode twice base64 encoded strings probably a nothingburger
English
@teortaxesTex Both models keep leaking their system prompt in CoT.
English
@teortaxesTex A lot of emphasis on agents in the description, sounds like Moonshot.
DS web repeats the question in CoT when I throw technical word salad at it every time, Hunter did it once out of 10 tries. Also it admitted to being a Chinese model out of the blue lol, I didn't even ask.
English
hunter does feel to have a similar vibe to what's on deepseek-Web, and sadly I can't say so far that it's stronger. Minor differences in reasoning style.
Healer seems to have worse general knowledge and not really frontier in multimodal understanding
AiBattle@AiBattle_
2 new Stealth models on OpenRouter Hunter Alpha: - "Hunter Alpha is a 1 Trillion parameter + 1M token context frontier intelligence model built for agentic use. It excels at long-horizon planning, complex reasoning, and sustained multi-step task execution, with the reliability and instruction-following precision that production agentic pipelines demand" Healer Alpha: - "Healer Alpha is a frontier omni-modal model with vision, hearing, reasoning, and action capabilities. It brings the full power of agentic intelligence into the real world: natively perceiving visual and audio inputs, reasoning across modalities, and executing complex multi-step tasks with precision and reliability"
English
@shalcker @teortaxesTex @nikola270179 Is occupying the shore necessary? The goal is to lower probability of ships being hit. Drone based surveillance and interception would be enough. That's maybe asking a lot from military but there are many people globally interested in getting the oil going, it's not just the US.
English
@teortaxesTex @nikola270179 Destroy all submarines, mines, guided torpedoes, and drone boats, as well as all capability to make new ones?
...drone boats that can be built in any garage and don't need large enterprise?
Or perhaps occupy entire Strait of Hormuz shore?
Two months sounds very optimistic.
English
The reality is that Iran can't back down because 1) chain of command is likely severed, the Supreme Leader is dead/incapacitated and can't order a ceasefire and 2) it's pointless, Israel will regroup and try to kill them anyway. I think we're watching an accelerated demolition.
Marko Jukic@mmjukic
Iran has explicitly said their plan is to force lasting concessions by inflicting maximum economic pain. That taboo is broken, it's happening. So Trump and Israel have to project strength and bomb Iran more just to maintain negotiating leverage. Escalating doom spiral.
English
@spandrell4 Risk premiums are going to suck and some of the fossil fuels are processed in fixed installations in gulf countries, but I doubt that it will get much worse than 2022 medium and long term.
English
@spandrell4 Iran has limited capability and many targets. That being said, so did Yemenis and the US failed to suppress them at Aden. The incentive is much stronger at Hormuz so most likely if not USN then some other party not interested in starving for Iran will ensure passage.
English
That would work if Israel didn't exist and actually run things. I truly don't see how Trump gets out of this.
Iran isn't opening Hormuz until they feel safe long term and Israel will never allow that.
Spencer Hakimian@SpencerHakimian
Let me predict the future for you. Oil Stays >$100 for a few more days. Trump panics as he sees his Poll Numbers collapse further. Trump makes some schizophrenic truth social post that the U.S. won this war. U.S. soldiers leave Iran. The Ayatollah’s Son just got a 40 year reign courtesy of the Red White & Blue. Trump & Hegseth bomb Cuba next.
English